Mitigating Class Imbalance in Vision-Based Anomaly Detection via NAUF Undersampling: A Case Study in Automated Quality Control for Hydrogen Storage Manufacturing

Addressing the issue of sample imbalance in vision-based anomaly detection tasks remains a critical focus. This work proposes a novel hybrid method that integrates learning-based multidimensional feature extraction with a Novel Adaptive Undersampling Framework (NAUF) for image-based anomaly detection, particularly when defect samples are extremely scarce. First, the proposed method extracts image features using the backbone of a pretrained deep learning network. Next, the undersampling technique NAUF is applied to these extracted features, balancing the highly imbalanced image samples while preserving essential information. Finally, the images are classified using multiple base classifiers within an ensemble learning framework. On a dataset of defects for the inner surfaces of high-pressure hydrogen storage tanks, the proposed method significantly improved the key performance metrics (AUCPRC) of KNN, Decision Tree, and Random Forest by 14.23%, 11.67%, and 7.52% respectively, while also increasing their inference speeds by 60.43%, 97.34%, and 86.36%, en-hancing the practical application value of these base classifiers. This work offers valuable insights and potential applications for improving quality control in automated manufacturing and other industrial settings where data imbalance is a common challenge.